1

AUSTRALIA’S FIRST DYNAMIC LAND COVER MAP

Lymburner L1, Tan P

1, Mueller N

1, Thackway R

2, Thankappan M

1, Islam A

1,

Randall L2 and Lewis A

1

1. National Earth Observation Group, Geoscience Australia

Symonston, ACT, Phone 6249 9587, Fax 6249 9938 leo.lymburner@ga.gov.au

2. Australian Bureau of Agricultural and Resource Economics–Bureau of Rural

Sciences, 18 Marcus Clark Street, Civic, ACT

Abstract

A Dynamic Land Cover Map (DLCM) for Australia has been developed to provide comprehensive and consistent land cover information to inform national and state level priority setting monitoring and reporting for sustainable farming practices, management of water resources, air quality, soil erosion, and forests. The relatively long term time series observations available in the DLCM can be used to assess the land cover dynamics of forests, woodlands, rangelands and cropping systems.

The DLCM method was developed using 16-day Enhanced Vegetation Index (EVI) composites collected at a 250-metre resolution using the Moderate Resolution Imaging Spectroradiometer (MODIS) for the period from 2000 to 2008. As required the DLCM method can be applied to the 2009-10 imagery. The MODIS time series for each pixel was analysed using an innovative technique that reduced each time series into 12 coefficients based on the statistical, phenological and seasonal characteristics of each pixel. These coefficients were then clustered and the resultant clusters labelled using Catchment Scale Land Use and the National Vegetation Information System datasets.

The classification scheme used to describe land cover categories in the DLCM conforms to the 2007 International Standards Organisation (ISO) land cover standard (19144-2). Land covers including all land and vegetation types are clustered into 34 ISO classes.

An accuracy assessment based on around 26,000 independent sites was used to validate the DLCM. As land cover classes are not generally clear-cut, but merge gradually from one to the other, a fuzzy-logic system was used to compare the 34 DLCM classes with the field data on a sliding scale. The match between the field data and the DLCM was exact in 30% of cases, very similar in 35% of cases, moderately similar in 10% of cases, somewhat similar in 18% of cases and a complete mismatch in 7% of cases. These results show a high degree of consistency between the DLCM and the site-based dataset.

2

Introduction

Land cover is the observed biophysical cover on the Earth’s surface. This includes various combinations of native vegetation types, soil types, exposed rocks and water bodies as well as anthropogenic elements, such as plantations, crop types and built environments. Complementary land information includes land use – the purpose to which land cover is committed, and land management practices – the approach taken to achieve a land use outcome.

High frequency national land cover images collected at a moderate resolution provide critical information needed for national and regional reporting and decision making in Australia (NLWRA 2007). This frequency enables researchers and managers to observe changes between successive image dates, which are referred to in this paper as dynamic land cover. Many agencies and research institutions collect and interpret a wide variety of land cover data. Increasingly, land cover data is required at national, state and regional levels, with annual or monthly updates. A review of the information required for reporting on national environmental indicators identified the need for consistent national information to report on change and trends in vegetation cover extent and condition against a baseline (Boland and Thackway 2008).

Investment in the development of land cover data and information is not coordinated, and Australia lacks consistent, complete, nationally agreed standards and protocols for land cover data. Atyeo and Thackway (2006) highlighted the deficiencies in Australia not having an agreed national approach for describing and mapping land cover. They noted the benefits of Australia adopting and implementing the Food and Agriculture Organisation’s Land Cover Classification (Di Gregorio and Jansen 2000) for translating and compiling existing land cover datasets. The lack of an agreed national approach to mapping and monitoring land cover change severely limits the usefulness of available land cover data.

Dynamic land cover information is needed at international, national, state and regional levels to inform a wide range of natural resources and emergency management purposes, particularly where information on change and trend are required.

At the international level land cover products are needed to assess the type and extent land degradation issues e.g. desertification, clearing of forested areas, food production and food security and natural disasters. Global land cover products have been produced by a number of international organisations and science communities - mostly based on coarse spatial resolution sensor data e.g. AVHRR. The international land cover products include AVHRR Land Cover (Wilson and Henderson-Sellers 1985), MODIS IGBP (Friedl et al., 2002), and Globcover (Bicheron et al., 2008).

At the national level land cover products are needed to evaluate the impact of investments and monitoring changes in the type and extent of resource condition including the management of ground cover, soil health, the condition of forests and other vegetation (NLWRA 2007, Boland and Thackway 2008). National land cover products have been developed to meet specific purposes

3

e.g. carbon accounting, management of rangelands and ground cover. Datasets are created to meet national and international reporting requirements and have either been compiled nationally from state datasets (Barson et al 1995) or created for particular themes e.g. forest carbon (Caccetta et al 2006)

At the State and Territory level land cover products are needed to meet legislated resource management requirements including soil, water, forest and native vegetation and threatened species. At this level datasets are created by government agencies at scales to inform planning and management (Barson et al 1995).

At the regional level land cover products are needed primarily to meet local planning, management and development purposes. At this level datasets are created by a wide range of suppliers with little commonalities between the approaches that are used.

Information on change in extent, trends over time and changes in resource condition are fundamental to a wide range of applications. However, experience shows that while many natural resource management and emergency planning and response applications would benefit from access to dynamic land cover products this is generally not available or accessible. In addition, land cover products developed for one level e.g. national, can have limited benefits when applied at a finer level e.g. state and regional. This is because classification schemes and data specifications are driven by the communities of interest at the broader level that are unlikely to meet the needs at the finer level. Added to these issues, is the quality and reliability of these products developed for different purposes e.g. internationally produced land cover maps often lack the levels of thematic detail and accuracy required for NRM applications.

There is a need in for Australia a coordinated land cover initiative. Through improved data collection and interpretation this would improve data quality and quantity for the same expenditure, enabling Australia to align activities to international processes and protocols including ISO standards on land cover and land cover change. This initiative would also be relevant at international through to regional levels for a wide range of natural resources and emergency management issues. Dynamic land cover information is needed to evaluate the impact of investments and resource condition monitoring, develop national accounts from forest carbon and water cover, and to inform legislated reporting requirements.

We outline the development and evaluation of a dynamic land cover map and time series dataset for Australia which can be used at regional and international levels for a wide range of natural resource management and emergency management applications.

Method

The DLCM was derived using a novel approach to time series analysis of Enhanced Vegetation Index (EVI) data. The MOD13Q1 product available at http://lpdaac.usgs.gov/modis/dataproducts.asp contains both EVI and the Normalised Difference Vegetation Index (NDVI). The MOD13Q1 product

4

composites 16 sequential days of TERRA MODIS satellite data into a single image.

The EVI is a standard product created from the composite data using the following band ratio of 250m and 500m spatial resolution bands:

LBCBCB

BBGEVI

32112

12

Where

G = gain factor = 2.5

B2 = MODIS 250m band 2: near infra-red

B1 = MODIS 250m band 1: visible red

B3 = MODIS 500m band 3: visible blue

C1 = aerosol resistance term = 6

C2 = aerosol resistance term = 7.5

L = canopy background adjustment = 1

The advantage of EVI over other vegetation indices is that it includes adjustment factors for atmospheric effects and is more sensitive to higher biomass targets, and in the MOD13Q1 product, the 16 day composition process provides a cloud-free product. The DLCM uses 8 years of EVI data from April 2000 through to April 2008. The data were downloaded from the LandProcesses Distributed Active Archive Center (LPDAAC), located at the U.S. Geological Survey (USGS) Earth Resources Observation and Science Center (EROS) http://LPDAAC.usgs.gov The data were then mosaicked and remapped using the MODIS Reprojection Tool (MRT) as described in Paget and King (2008)

The novel approach involved characterising each EVI time series using a robust suite of statistical, phenological and seasonal factors.

Statistical o Average Annual Mean o Average Annual Maximum o Average Annual Minimum o Average Annual Standard Deviation o Global Minimum o Ratio of the Average Annual Maximum to the Global Maximum

Phenological o Average annual rate of rise o Average annual rate of drop o Flatness

Seasonal o Timing of peak EVI value o Standard deviation in the timing of the peak o Length of the growing cycle

5

These factors are referred to as time series coefficients. The resultant time series coefficients were then ‘clustered’ into groups of pixels and labelled as land cover units. The clusters were initially labelled using the Native Vegetation Information System and the Catchment Scale Land Use datasets. These initial labels were reviewed and refined by:

An expert panel at a national workshop in Alice Springs in September 2009,

Comparison with the Fraction of Photosynthetically Active Radiation (FPAR) dataset (Donohue et. al., 2008),

Comparison with the Foliage Projective Cover (FPC) dataset produced by the Statewide Land and Tree Survey (SLATS) team from the Queensland Department of Environment and Resource Management, and

Stratification based on the Interim Biogeographic Regionalisation of Australia (IBRA) dataset (Thackway and Cresswell 1995).

The accuracy of the DLCMv1 was assessed by comparing it with 25,817 field survey data points provided by the State and Territory vegetation and land management agencies and by the Australian Bureau of Agriculture and Resource Economics (ABARE-BRS). Table 1 presents a list of sources of field data points used in the accuracy assessment of DLCM and Figure 2 Shows the spatial distribution of these points.

Table 1 List of sources of field data points used in the accuracy assessment of DLCM

Data Source Classification Type Number of Points

ABARE-BRS Managed Land Covers 1,215

Australian Capital Territory Department of Parks Conservation and Lands

Vegetation 19

Northern Territory Department of Environment, Climate Change and Water

Vegetation 1,859

New South Wales Department of Environment, Climate Change and Water

Vegetation and Managed Land Cover

882

Queensland Herbarium Vegetation 16,992

South Australian Department of Environment and Heritage

Vegetation and Managed Land Cover

3,450

Tasmanian Department of Primary Industries, Parks, Water and Environment

Vegetation 443

Victorian Department of Sustainability Vegetation 20

6

and Environment

WA Department of Agriculture and Food

Vegetation 1,389

A fuzzy-logic approach was used to access the accuracy of the DLCM using the method developed by Lowry et al. (2007). The system employs a relative similarity score to compare situations where the classes are not mutually exclusive and large overlaps may occur, as is intrinsic to land cover. This necessary because in comparing the DLCM to the field dataset there are many instances where there was no direct one-to-one match between the DLCM class and the different field survey pro forma used to collect the datasets outlined in Table 1. Given the diversity of classification schemas used by different field survey teams, and the large number of classes (>70) in some schema, the analysis contained in this report only describes the Users Accuracy rather than Producers Accuracy (Congalton and Green 2002). To account for diversity of classification schemas used by the field survey teams and to account for the semantic differences between the class names used by the survey teams and the ISO standard (i.e. the classification schema used by the DLCM) the fuzzy approach was modified to allow classes to be labelled as semantically identical between the DLCM and a compared dataset. A Relative Similarity Score (RSS) matrix was created for each comparison, ranking an exact semantic match between classes as a value of 5, with decreasing similarity gradually lowering the score to a no-similarity value of 1

The classes of the sliding scale are:

complete match such as trees open vs trees open (score =5),

very similar such as trees open vs trees scattered (score =4),

moderately similar as in trees open vs shrubs open (score =3),

somewhat similar as in trees open vs shrubs closed (score =2), and

complete mismatch as in trees open vs irrigated graminoids (score =1).

For example the West Australian field survey data contains points classified as Gascoyne bluebush. The West Australian flora database FloraBase http://florabase.dec.wa.gov.au/ identifies Gascoyne bluebush as a chenopod making it semantically identical to the chenopod classes contained within the DLCM. However because the structure (open or sparse) is not recorded in the field survey, Gascoyne blue bush is rated as a 4 (very similar) to both open chenopod and sparse chenopod, but not a 5 (direct match for either). Similarly Notophyll Vine Forest in the Queensland field survey database was taken as a complete match (5) with closed forest. In instances where the class description i.e. ‘Riverine mixed’ did not contain sufficient information to clearly identify a comparable class within the DLCM, the sites from that class were omitted from the analysis.

The fuzzy match concept also provides the capacity to characterise the gradually changing nature of woody vegetation structure. For example there

7

are areas where a dense woodland is indistinguishable from open forest, in the sense that the difference between the two is a subtle (and in many cases gradual) difference in leaf area).

Results and Discussion

Figure 1 (a and b) show the type and extent of DLCM classes. Figure 1a shows that grasslands and shrublands comprise almost half the total number of DLCM classes (i.e. 8 and 7 classes respectively). This compares to only four classes that are used to describe the forest and woodland cover types. Agricultural and pasture management systems are described by four and two DLCM classes respectively.

Table 2 summarises the 34 DLCM cover types in Figure 1a into nine broad cover classes. Consistent with other published information, Australia’s landscapes are dominated by grasslands and shrublands, accounting for 37% and 21% respectively, being located mainly in the rangelands. Forest and woodland cover types account for 32% of the land surface and are characteristically found around the continent’s northern, eastern and southern margins. This broad woody cover class includes open woodland, savannas, open and closed forests and plantations. Agricultural landscapes and pasture management systems (rainfed and irrigated) including agricultural systems and pasture account for 5% and 4% respectively. These cover types are found in the south western and south eastern hinterlands and coastal margins of eastern Australia.

Table 2. Number, and area of broad land cover types presented in Figure 1b.

Theme Area (ha) Per cent of Australia

Not labelled 15,838 0.0% Bare and mining 31,063 0.0% Sedge and alpine 89,513 0.0% Aquatic and forb 338,656 0.0% Water and salt lakes 11,351,744 1.5% Pastures 29,866,938 3.9% Agriculture 35,527,894 4.6% Shrublands 159,451,706 20.7% Woodlands and forests

1 247,185,688 32.2%

Grasslands 284,944,738 37.1%

Total 768,803,775 100.0%

1 This area includes the 149 mill ha of Australia’s forests as defined in the Australia’s Agriculture, Fisheries and

Forestry at a glance 2010 (http://www.daff.gov.au/__data/assets/pdf_file/0003/1539831/aag-2010.pdf).. The largest

proportion of this 247 mill ha includes the class ‘trees – sparse’ i.e. open woodlands, comprising 129 mill ha.

8

Figure 1a The Dynamic Land Cover Map version 1.

Figure 1b the legend for Figure 1a.

NB: Land cover classes including trees, shrubs and graminoids are defined in the International Standards

Organisation (ISO) land cover classification.

9

Proportional Accuracy Assessment

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

Water and Salt

Lakes (271)

Grasslands (2504) Shrublands (1677) Woodland and

Forests (18983)

Pastures (1337) Agricultural

Systems (983)

Broad Land Cover Types

Complete Mismatch

Somewhat Similar

Moderately Similar

Very Similar

Complete Match

Figure 3. Thematic breakdown of the accuracy assessment for broad land cover types

Applying the relative similarity score approach to all 25,817 field survey points and calculating the accuracy of the DLCMv1 against these scores yielded a complete match of 30%, a very similar score of 35%, a moderately similar score of 10%, somewhat similar score of 18% and a complete mismatch of 7%. The accuracy for each broad land cover type is described in Figure 3, land cover types that had less than 50 validation points have been omitted from the figure. The reasons for differences between the DLCM and the field data are discussed below.

Thematic emphasis of the field survey data

The field survey data used to provide the 25,817 accuracy assessment were collected by different agencies for different purposes. Some of the surveys have a botanical emphasis, others have a vegetation structural emphasis, and others have a rangelands or land use emphasis. As a consequence of this, some land cover classes are under represented from a validation point of view in some regions. For example, the West Australian field survey was rangeland specific, so the forest classes are not present in that dataset. None of the field validation datasets focussed on irrigated cropping, irrigated pastures, sugar (either irrigated or rainfed), water, or mining. These classes have very few validation points (<10 per class with the exception of the water classes), and so it is difficult to draw any definitive conclusions about the accuracy of the DLCM for these land cover classes. Different field survey data sets would be required to properly assess the accuracy of these classes.

Semantic differences in labelling

The different thematic emphases of the different datasets meant that the field survey data labels were often not directly comparable with the DLCM ISO

10

standard classes as they are derived from sources using labelling conventions particular to their own needs. For example rainfed pasture is an ISO class, but the botanical emphasis of many of the field datasets does not provide an equivalent class to enable direct accuracy assessment. Similarly, the botanical emphasis of many of the vegetation survey data means that many classes are genus or species based rather than common name based; Astrebla sp. for example needs to be converted to Tussock Grassland. This is easy to address for the major species, but the level of detail can confound this process. Many classes describe the genus and/or species but not the structural component, so it is not possible to identify a complete match. For example if the density of the Astrebla sp. grassland isn’t specified, then it can be assigned a similarity rating of 4 with all of the Tussock grassland classes, but it is not possible to identify a complete match.

Land Cover classes that are poorly characterised by EVI

Some classes within the ISO class nomenclature are not well characterised by EVI. Classes such as extractive site, which is the ISO land cover term for quarries and mines, are characterised by very low to zero vegetation and are therefore not clearly linked to greenness, so the accuracy of these classes is likely to be poor. Other land cover with little vegetation such as open tussock grassland and sparse hummock grassland also have relatively low ‘complete match’ or close match (complete + very similar), due to the fact that areas with very sparse vegetation, that responds to very infrequent and randomly timed rainfall events, are difficult to distinguish using EVI. However these very low density grassland and shrubland areas are rarely a complete mismatch; they are more often mislabelled as another very low density vegetation type that fall into the moderately or somewhat similar categories.

Systematic Class vs Canopy Continuum

Within the denser canopy classes (trees-closed - trees sparse) and (shrubs closed – shrubs sparse) there is a low to moderate (12%-49%) complete match rate but a moderate to high (33%-85%) very close (complete + very similar) match rate. This reflects the reality that vegetation structural classes are a series of systematic breaks in a continuous change in canopy closure and height, i.e. there are some instances where dense woodland will look almost indistinguishable from open forest. Similarly there are instances where a tall open forest is very similar to a closed forest. Which may be a function of site condition and/or disturbance history Evaluation of the individual state and territory datasets reveals a systematic over-estimation of canopy closure class within the DLCM, particularly with open forest in the field survey data being classified as closed forest by DLCM. These errors of overestimation or ‘producers error’ sensu Congalton (1991) are impossible to report on in a consistent fashion because of the sheer diversity of class names used in the field survey datasets.

Scale and perspective of observation of the field survey data

There are locations where due to the resolution of the DLCM, where there are likely complete mismatches with the field observations. For example some

11

areas where sharp environmental gradients are observed, such as narrow riparian strips, narrow strips of coastal vegetation and small patches of remnant vegetation in areas that have been cleared and converted to intensive land uses. Other mismatches occur where the information in the training data for an area (e.g. NVIS) didn’t fully characterise the range of greenness dynamics (particularly mining and bare areas). Some of the field survey locations were located in spatial settings that are not consistent with comparison with 250m x 250m pixels. Several vegetation surveys are located in narrow riparian strips, small patches of remnant vegetation, or in coastal areas where there are narrow strips and bands of vegetation communities associated with fringing mangrove forests and dune-swale complexes. These sites were not omitted from the analysis because it is important to recognise that there are areas that contain sharp environmental gradients where the 250 metre DLCM will not accurately represent the land cover present on the ground.

Descriptions of the field sites, particularly those derived from botanical and vegetation structural surveys tend to focus on the tallest stratum, with variable levels of detail used to describe the lower stratum. This is consistent with the predominant botanical and structural nomenclatures, and with a horizontal or ‘looking out’ perspective. Earth observation data on the other hand has a vertical or ‘looking down’ perspective, so the influence of the lower stratum on the EVI signal increases rapidly as the upper stratum becomes more open. A field survey site described as low open woodland with no description of the lower stratum may be a low open woodland with open tussock grassland understorey; in this scenario, a DLCM classification of open tussock grassland would be rated as somewhat similar (because there is no description of the lower stratum) but the DLCM is very similar. As a consequence the error assessment may under-estimate the true accuracy of the DLCM where the field survey lacks a description of the lower stratum.

Applications

The DLCM can be used as input into moderate to coarse scale models of:

Groundwater recharge and discharge

Climate

Wind and Water erosion risk

Evapotranspiration

Carbon dynamics

Land surface processes

Inundation

The DLCM (Figure 1 a and b) in combination with the trend in annual EVI (e.g. Figure 3) can provide insight into the response of land cover to a wide variety of drivers, both natural and anthropogenic. This provides natural resource managers with the capacity to identify emerging patterns of land cover change, and provides a broad spatial and historical context within which to interpret that land cover change. This can be combined with ancillary information to assess

12

what, if any, on-ground or policy interventions are required to mitigate the emerging behaviour.

Cropping in northern WA wheat belt

0

0.1

0.2

0.3

0.4

0.5

0.6

0 23 46 69 92 115 138 161 184

Time

En

han

ced

Veg

eta

tio

n In

dex

Crop

2000 2001 2002 2003 2004 2005 2006 2007

Figure 4. EVI time series for one pixel in a cropping region in WA, the green line represents

the trend in the annual mean EVI values

Our assessment of the DLCM showed that these information can also be used as a framework for analysing other coarse resolution datasets such as 500m fractional cover data (Guerschman et al. 2009) and 1km fraction of photosynthetically absorbed radiation data (fPAR)(Donohue et al. 2009), thereby providing further insight into land cover change processes. Using the DLCM as a framework to interpret EVI, and fractional cover provides NRM decision makers with additional knowledge on which to form evidence-based policy.

Conclusions

The DLCM is a consistent, complete land cover and time series product for all of Australia. It provides a national baseline for land cover analyses and management. The flexibility of this new dataset offers the potential for researchers to investigate whether it can be used to meet some of Australia’s national environmental reporting needs. The daily frequency of MODIS observations provides the DLCM with the ability to be generated at regular intervals aligned with statutory needs, making seasonal and even monthly updates possible. Consequently the DLCM could be used to provide consistent national, state and regional land cover change information. Derivative products from the DLCM such as EVI trend analyses have the potential to provide large scale comparisons of the performance of significant environmental features such as mangroves or alpine grasslands in response to natural and/or

13

anthropogenic influences. Used as a platform to underpin detailed land cover projects, the DLCM provides a framework to align independent analyses to a common point of comparison.

The development of the DLCM highlights the usefulness of time series analyses in developing a consistent land cover classification and change products. The composite MODIS vegetation index approach helps mitigate the effects of clouds and other interferences (such as data errors) that can make standard satellite image approaches unworkable. The dynamic vegetation index approach provides a consistent reference for every class, hence classes are inter-comparable and may be related to biophysical variables such as fPAR and fractional cover.

Acknowledgements

The creation of the Dynamic Land Cover Map was made possible by contributions from a number of organisations across Australia. These contributions included the provision of data for the initial creation and labelling of the classes, comparative data to revise and validate the land cover classes, detailed validation work and participation in the Dynamic Land Cover Map workshop. We thank the following organisations for valuable contributions to the development of the Dynamic Land Cover Map:

Australian Collaborative Land Use Mapping Program (ACLUMP),

Australian Collaborative Rangelands Information System (ACRIS),

Commonwealth Scientific and Industrial Research Organisation (CSIRO),

Executive Steering Committee for Australian Vegetation Information (ESCAVI),

National Committee for Land Use and Management Information (NCLUMI),

Queensland Statewide Landcover And Tree Study (SLATS),

State and Territory Government Departments of New South Wales, The Australian Capital Territory, Northern Territory, Queensland, South Australia, Tasmania, Victoria and Western Australia.

References

Atyeo C. and R. Thackway (2006). Classifying Australian land cover. Bureau of Rural Sciences, Canberra. Pp 27.

Bicheron P., Defourny P., Brockmann C., Schouten L., Vancutsem C., Huc M., Bontemps S., Leroy M., Achard F., Herold M., Ranera F., Arino O., (2008). GLOBCOVER - Products Description and Validation Report. Toulouse, France. Pp 47.

Boland, C and R. Thackway (2008). Analysis of the requirements of National Coordinating Committee’s for vegetation-related and landcover information. (October, 2008). Bureau of Rural Sciences, Canberra.

Congalton, R. (1991) A review of assessing the accuracy of classifications of remotely sensed data Remote Sensing of Environment 37, (1), 35-46

14

Donohue, Randall J.; Mcvicar, Tim R.; Roderick, Michael L.(2009) ‘Climate-related trends in Australian vegetation cover as inferred from satellite observations. 1981-2006’ Global Change Biology, 15, (4), 2009, 1025-1039

Friedl M.A, McIver D.K. , Hodges J.C.F , Zhang X.Y ,. Muchoney D , Strahler A.H. ,Woodcock C.E.,. Gopal S, Schneider A, Cooper A, Baccini A, Gao F, and Schaaf C. (2002) ‘Global land cover mapping from MODIS: algorithms and early results’ Remote Sensing of Environment 83 (2002) 287–302NLWRA (National Land & Water Resources Audit) (2007). Australian land cover mapping. Proceedings of a workshop to discuss interest in land cover mapping. 24 July 2007. National Land & Water Resources Audit. Canberra.

Guerschman, J. P., Hill, M. J., Renzullo, L. J., Barrett, D. J., Marks A. S., Botha, E. J., (2009). Estimating fractional cover of photosynthetic vegetation, non-photosynthetic vegetation and bare soil in the Australian tropical savannah region upscaling Hyperion and MODIS sensors. Remote Sensing of Environment, 113, 928-945

Lowry, J., Langs, L, Ramsey, R., Kirby, J., Schultz, K. (2007). A matrix-based approach to fuzzy set accuracy assessment. RS/GIS Lab White Paper, Remote Sensing/GIS Laboratory, College of Natural Resources, Utah State University.

Paget MJ and King EA (2008) MODIS Land data sets for the Australian region. CSIRO Marine and Atmpospheric Research Internal Report No. 004.

Thackway, R. and Cresswell, I.D. 1995. An interim biogeographic regionalisation for Australia: a framework for setting priorities in the National Reserves System Cooperative Program, Version 4.0. Australian Nature Conservation Agency, Canberra.

Wilson, M., & Henderson-Sellers, A. (1985). A global archive of land cover and soils data for use in general circulation climate models. Journal ofClimatology, 5, 119–143.